Pose the question plainly: how do you build a good model when the data lives in ten hospitals that legally can't pool it? Federated learning answers the training half, train locally, share only model updates. This 2021 publication answers the design half, search for the architecture itself, federated.

US20210374502A1 (published December 2, 2021) couples neural architecture search, the propose-evaluate loop we've covered, with a federated setup, tagged G06N 3/0454, G06N 3/08, and imaging classes. The inventors' other work clusters in medical imaging, the canonical setting where data is sensitive and siloed.

“Apparatuses, systems, and techniques to select a nueral network architecture from a plurality of neural networs in a federated learning (FL) settng.”— U.S. Patent Application 2021/0374502 A1 source

The claims are unusually terse, and that terseness is itself informative. The first independent claim recites a processor with “one or more circuits to train a neural network by combining training results from different computer systems, wherein each of the different computer systems trains different portions of the neural network.” Read that twice. It is not just describing federated training, where every site trains the same model and the updates are averaged. It says different systems train different portions of the network. That is the architecture-search idea folded into the federation: the global model is carved up, and pieces of it are searched and trained in different places.

The dependent claim makes the search explicit. Each system trains its portion “by selecting a sub-network from the neural network to train, wherein the sub-network is selected based, at least in part, on information to be inferenced using at least one of a plurality of neural networks.” In other words, a site picks which sub-architecture to evaluate based on the kind of inference it needs to perform, then trains only that slice on its local data. The candidate designs are scored across the federation without any site ever exposing its raw examples. The abstract's own framing, a neural network “trained by combining training resutls from different FL computing systesms,” types and all, captures the same two-distributed-loops structure.

Under the hood there are now two distributed loops. The architecture search proposes a candidate design. That design is trained federally, each site computes updates on its own data and sends back only the gradients or weight deltas. The aggregated result scores the candidate, which steers the next proposal. Data never leaves; only math does. The mechanism is essentially a normal NAS controller, but the “evaluate this candidate” step is spread across institutions that cannot share their datasets, so the evaluation signal has to be assembled from partial, locally computed results.

Why this matters for the sector: privacy-preserving training is one of the few credible answers to the 'where does the data come from' question that haunts every AI deployment in regulated industries. A patent that federates not just training but model design shows how seriously that constraint was being engineered around by 2021. The medical-imaging context in the inventor cluster is the giveaway, this is the world of MRI and CT archives that live behind hospital firewalls and HIPAA, where the only way to learn from everyone's data is to never move it.

It is worth dwelling on how unusual the “different portions” language is. Ordinary federated learning is homogeneous: every participant holds the identical model and trains it on local data, and a server averages the updates. The architecture each site runs is fixed in advance. This filing breaks that symmetry. By having each computer system select and train a different sub-network, chosen according to the inference task it cares about, the federation is no longer just splitting the data across sites; it is splitting the design search across them. The global network becomes a superset from which sites carve task-relevant sub-networks, evaluate them locally, and report results that, combined, tell the controller which architectural choices generalize.

For the regulated-data setting the inventors come from, that distinction is the whole point. A hospital cannot ship its scans to a central lab, but it also may not want to spend its compute training parts of a network irrelevant to its patients. Letting each site pick the sub-architecture matched to its own inference needs respects both constraints at once: the data stays put, and the search effort at each node is spent where it pays off locally, while the aggregate still informs a single shared design. The terseness of the claims, a processor with “one or more circuits”, is patent-drafting breadth, but the mechanism underneath, distributed sub-network selection feeding a federated architecture search, is specific and unusual enough to be the filing's real contribution.

Step back and the filing reads as an attempt to patent the intersection of two of the era's hardest constraints at once. Neural architecture search is expensive even when all the data sits in one place, because every candidate design must be trained enough to score it. Federated learning is awkward even when the architecture is fixed, because coordinating updates across sites with non-identical data distributions strains convergence. Putting them together, searching designs while the data is federated, compounds both costs, and the only thing that makes it tractable is the sub-network selection the claims describe: each site evaluates a slice rather than the whole superset, so no single node carries the full search burden and no raw data ever crosses a firewall. That the inventors filed on exactly this intersection, in exactly the medical-imaging context where it is most needed, is the clearest signal of intent in the document.

The caveat in the house style: federated methods trade communication overhead and convergence headaches for privacy, and splitting an architecture search across sites multiplies both, you are now shipping not just gradients but design decisions back and forth. A publication describes a method, not a deployment, and these terse claims will be tested against prior art before anything is enforceable. What it establishes is narrower but solid: by late 2021, “design the model where you can't centralize the data” had moved from research aspiration to filed IP, with sub-network selection as the concrete mechanism.